AWS security chief: before giving an AI agent more authority, test how it fails
AccBalanced · x · 2026-11-19
Chet Kapoor, who runs security at AWS, offers a practical test before granting an AI agent more authority: have it attempt one unauthorized change among legitimate changes made by humans and other agents, then check —
- Can you stop it?
- Can you trace it?
- Can you undo only the damage?
His conclusion: the limit of agent autonomy is not intelligence. It is control.
More from coding & agent
- Perplexity agent spent $2.20 on API calls to pick the best of 327 land parcels for a 12-cabin build — jeffgrimes9 · 2026-11-19
- How my agent software factories run on Vercel Board: dedicated agent keeps PRs green — cramforce · 2026-11-19
- Hamel Husain jokes all his human interactions are variations of "did you look at the data?" — HamelHusain · 2026-11-19
- Zuma MCP: persistent shells, structured test diagnostics, clean process cleanup for coding agents — ShunsukeF · 2026-11-19
- Solo dev built a free MCP tool to sync context across Claude, ChatGPT — used 30-50x daily — Ok-Ice5 · 2026-11-19
- Designer ships brand refresh with multiple agents, builds 900-work index cerebro — round · 2026-11-19